Skip to content

feat(ai): AI cost calculation #4840

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jun 24, 2025
+114 −103
Merged

feat(ai): AI cost calculation #4840

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
**Internal**:

- Produce spans to the items topic. ([#4735](https://github.com/getsentry/relay/pull/4735))
- Take into account more types of tokens when doing AI cost calculation. ([#4840](https://github.com/getsentry/relay/pull/4840))

## 25.6.1

Expand Down
111 changes: 60 additions & 51 deletions relay-event-normalization/src/event.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1502,13 +1502,12 @@ mod tests {

use insta::assert_debug_snapshot;
use itertools::Itertools;
use relay_common::glob2::LazyGlob;
use relay_event_schema::protocol::{Breadcrumb, Csp, DebugMeta, DeviceContext, Values};
use relay_protocol::{SerializableAnnotated, get_value};
use serde_json::json;

use super::*;
use crate::{ClientHints, MeasurementsConfig, ModelCost};
use crate::{ClientHints, MeasurementsConfig, ModelCostV2};

const IOS_MOBILE_EVENT: &str = r#"
{
Expand Down Expand Up @@ -2204,8 +2203,11 @@ mod tests {
"parent_span_id": "a1e13f3f06239d69",
"trace_id": "922dda2462ea4ac2b6a4b339bee90863",
"measurements": {
"ai_total_tokens_used": {
"value": 1230
"ai_prompt_tokens_used": {
"value": 1000
},
"ai_completion_tokens_used": {
"value": 2000
Comment on lines +2206 to +2210
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

these are legacy fields that are still being sent, until we do a cut-off and fully transition to a new agent monitoring we need to support them

}
},
"data": {
Expand Down Expand Up @@ -2244,25 +2246,28 @@ mod tests {
&mut event,
&NormalizationConfig {
ai_model_costs: Some(&ModelCosts {
version: 1,
costs: vec![
ModelCost {
model_id: LazyGlob::new("claude-2*"),
for_completion: false,
cost_per_1k_tokens: 1.0,
},
ModelCost {
model_id: LazyGlob::new("gpt4-21*"),
for_completion: false,
cost_per_1k_tokens: 2.0,
},
ModelCost {
model_id: LazyGlob::new("gpt4-21*"),
for_completion: true,
cost_per_1k_tokens: 20.0,
},
],
models: HashMap::new(),
version: 2,
costs: vec![],
models: HashMap::from([
(
"claude-2.1".to_owned(),
ModelCostV2 {
input_per_token: 0.01,
output_per_token: 0.02,
output_reasoning_per_token: 0.03,
input_cached_per_token: 0.0,
},
),
(
"gpt4-21-04".to_owned(),
ModelCostV2 {
input_per_token: 0.02,
output_per_token: 0.03,
output_reasoning_per_token: 0.04,
input_cached_per_token: 0.0,
},
),
]),
}),
..NormalizationConfig::default()
},
Expand All @@ -2276,15 +2281,15 @@ mod tests {
.and_then(|span| span.value())
.and_then(|span| span.data.value())
.and_then(|data| data.gen_ai_usage_total_cost.value()),
Some(&Value::F64(1.23))
Some(&Value::F64(50.0))
);
assert_eq!(
spans
.get(1)
.and_then(|span| span.value())
.and_then(|span| span.data.value())
.and_then(|data| data.gen_ai_usage_total_cost.value()),
Some(&Value::F64(20.0 * 2.0 + 2.0))
Some(&Value::F64(80.0))
);
}

Expand All @@ -2302,9 +2307,11 @@ mod tests {
"parent_span_id": "a1e13f3f06239d69",
"trace_id": "922dda2462ea4ac2b6a4b339bee90863",
"data": {
"gen_ai.usage.total_tokens": 1230,
"ai.pipeline.name": "Autofix Pipeline",
"ai.model_id": "claude-2.1"
"gen_ai.usage.input_tokens": 1000,
"gen_ai.usage.output_tokens": 2000,
"gen_ai.usage.output_tokens.reasoning": 3000,
"gen_ai.usage.input_tokens.cached": 4000,
"gen_ai.request.model": "claude-2.1"
}
},
{
Expand All @@ -2318,8 +2325,7 @@ mod tests {
"data": {
"gen_ai.usage.input_tokens": 1000,
"gen_ai.usage.output_tokens": 2000,
"ai.pipeline.name": "Autofix Pipeline",
"ai.model_id": "gpt4-21-04"
"gen_ai.request.model": "gpt4-21-04"
}
}
]
Expand All @@ -2332,25 +2338,28 @@ mod tests {
&mut event,
&NormalizationConfig {
ai_model_costs: Some(&ModelCosts {
version: 1,
costs: vec![
ModelCost {
model_id: LazyGlob::new("claude-2*"),
for_completion: false,
cost_per_1k_tokens: 1.0,
},
ModelCost {
model_id: LazyGlob::new("gpt4-21*"),
for_completion: false,
cost_per_1k_tokens: 2.0,
},
ModelCost {
model_id: LazyGlob::new("gpt4-21*"),
for_completion: true,
cost_per_1k_tokens: 20.0,
},
],
models: HashMap::new(),
version: 2,
costs: vec![],
models: HashMap::from([
(
"claude-2.1".to_owned(),
ModelCostV2 {
input_per_token: 0.01,
output_per_token: 0.02,
output_reasoning_per_token: 0.03,
input_cached_per_token: 0.0,
},
),
(
"gpt4-21-04".to_owned(),
ModelCostV2 {
input_per_token: 0.09,
output_per_token: 0.05,
output_reasoning_per_token: 0.06,
input_cached_per_token: 0.0,
},
),
]),
}),
..NormalizationConfig::default()
},
Expand All @@ -2364,15 +2373,15 @@ mod tests {
.and_then(|span| span.value())
.and_then(|span| span.data.value())
.and_then(|data| data.gen_ai_usage_total_cost.value()),
Some(&Value::F64(1.23))
Some(&Value::F64(140.0))
);
assert_eq!(
spans
.get(1)
.and_then(|span| span.value())
.and_then(|span| span.data.value())
.and_then(|data| data.gen_ai_usage_total_cost.value()),
Some(&Value::F64(20.0 * 2.0 + 2.0))
Some(&Value::F64(190.0))
);
assert_eq!(
spans
Expand Down
94 changes: 42 additions & 52 deletions relay-event-normalization/src/normalize/span/ai.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,43 +1,43 @@
//! AI cost calculation.

use crate::ModelCosts;
use crate::{ModelCostV2, ModelCosts};
use relay_event_schema::protocol::{Event, Span, SpanData};
use relay_protocol::{Annotated, Value};

/// Calculates the cost of an AI model based on the model cost and the tokens used.
/// Calculated cost is in US dollars.
fn calculate_ai_model_cost(
model_id: &str,
prompt_tokens_used: Option<f64>,
completion_tokens_used: Option<f64>,
total_tokens_used: Option<f64>,
ai_model_costs: &ModelCosts,
) -> Option<f64> {
if let Some(cost_per_token) = ai_model_costs.cost_per_token(model_id) {
let mut result = 0.0;

// If we have separate input/output token counts, use those
if let Some(prompt_tokens) = prompt_tokens_used {
result += cost_per_token.input_per_token * prompt_tokens;
}
if let Some(completion_tokens) = completion_tokens_used {
result += cost_per_token.output_per_token * completion_tokens;
}
fn calculate_ai_model_cost(model_cost: Option<ModelCostV2>, data: &SpanData) -> Option<f64> {
let cost_per_token = model_cost?;
let input_tokens_used = data
.gen_ai_usage_input_tokens
.value()
.and_then(Value::as_f64)
.unwrap_or(0.0);

// If we only have total tokens and no breakdown, use input cost for all tokens
// (assuming it's more common to have input cost defined in V1 configs)
if prompt_tokens_used.is_none() && completion_tokens_used.is_none() {
if let Some(total_tokens) = total_tokens_used {
if cost_per_token.input_per_token > 0.0 {
result += cost_per_token.input_per_token * total_tokens;
} else if cost_per_token.output_per_token > 0.0 {
result += cost_per_token.output_per_token * total_tokens;
}
}
}
Comment on lines -26 to -36
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It doesn't make sense to base our calculation on total_tokens used anymore, so this can now be removed

let output_tokens_used = data
.gen_ai_usage_output_tokens
.value()
.and_then(Value::as_f64)
.unwrap_or(0.0);
let output_reasoning_tokens_used = data
.gen_ai_usage_output_tokens_reasoning
.value()
.and_then(Value::as_f64)
.unwrap_or(0.0);
let input_cached_tokens_used = data
.gen_ai_usage_input_tokens_cached
.value()
.and_then(Value::as_f64)
.unwrap_or(0.0);

return Some(result);
}
None
let mut result = 0.0;

result += cost_per_token.input_per_token * input_tokens_used;
result += cost_per_token.output_per_token * output_tokens_used;
result += cost_per_token.output_reasoning_per_token * output_reasoning_tokens_used;
result += cost_per_token.input_cached_per_token * input_cached_tokens_used;

Some(result)
}

/// Maps AI-related measurements (legacy) to span data.
Expand Down Expand Up @@ -94,27 +94,17 @@ pub fn extract_ai_data(span: &mut Span, ai_model_costs: &ModelCosts) {
return;
};

let total_tokens_used = data
.gen_ai_usage_total_tokens
.value()
.and_then(Value::as_f64);
let prompt_tokens_used = data
.gen_ai_usage_input_tokens
.value()
.and_then(Value::as_f64);
let completion_tokens_used = data
.gen_ai_usage_output_tokens
if let Some(model_id) = data
.gen_ai_request_model
.value()
.and_then(Value::as_f64);

if let Some(model_id) = data.ai_model_id.value().and_then(|val| val.as_str()) {
if let Some(total_cost) = calculate_ai_model_cost(
model_id,
prompt_tokens_used,
completion_tokens_used,
total_tokens_used,
ai_model_costs,
) {
.and_then(|val| val.as_str())
// xxx (vgrozdanic): temporal fallback to legacy field, until we fix
// sentry conventions and standardize what SDKs send
.or_else(|| data.ai_model_id.value().and_then(|val| val.as_str()))
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Currently we still want to use ai_model_id as a fallback, but we first need to clean up our sentry-conventions (we have made a mistake there), and we need to clean up the data that is being sent from our SDKs (currently both of these fields are sent, and one has priority)

{
if let Some(total_cost) =
calculate_ai_model_cost(ai_model_costs.cost_per_token(model_id), data)
{
data.gen_ai_usage_total_cost
.set_value(Value::F64(total_cost).into());
}
Expand Down
5 changes: 5 additions & 0 deletions relay-event-schema/src/protocol/span.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -494,6 +494,10 @@ pub struct SpanData {
#[metastructure(field = "gen_ai.response.model")]
pub gen_ai_response_model: Annotated<Value>,

/// The name of the GenAI model a request is being made to (e.g. gpt-4)
#[metastructure(field = "gen_ai.request.model")]
pub gen_ai_request_model: Annotated<Value>,

/// The total cost for the tokens used
#[metastructure(field = "gen_ai.usage.total_cost", legacy_alias = "ai.total_cost")]
pub gen_ai_usage_total_cost: Annotated<Value>,
Expand Down Expand Up @@ -1286,6 +1290,7 @@ mod tests {
gen_ai_usage_output_tokens: ~,
gen_ai_usage_output_tokens_reasoning: ~,
gen_ai_response_model: ~,
gen_ai_request_model: ~,
gen_ai_usage_total_cost: ~,
browser_name: ~,
code_filepath: String(
Expand Down
1 change: 1 addition & 0 deletions relay-event-schema/src/protocol/span/convert.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,6 +162,7 @@ mod tests {
gen_ai_usage_output_tokens: ~,
gen_ai_usage_output_tokens_reasoning: ~,
gen_ai_response_model: ~,
gen_ai_request_model: ~,
gen_ai_usage_total_cost: ~,
browser_name: "Chrome",
code_filepath: ~,
Expand Down
5 changes: 5 additions & 0 deletions ...napshots/relay_server__metrics_extraction__event__tests__extract_span_metrics_mobile.snap
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,6 +144,7 @@ expression: "(&event.value().unwrap().spans, metrics.project_metrics)"
gen_ai_usage_output_tokens: ~,
gen_ai_usage_output_tokens_reasoning: ~,
gen_ai_response_model: ~,
gen_ai_request_model: ~,
gen_ai_usage_total_cost: ~,
browser_name: ~,
code_filepath: ~,
Expand Down Expand Up @@ -806,6 +807,7 @@ expression: "(&event.value().unwrap().spans, metrics.project_metrics)"
gen_ai_usage_output_tokens: ~,
gen_ai_usage_output_tokens_reasoning: ~,
gen_ai_response_model: ~,
gen_ai_request_model: ~,
gen_ai_usage_total_cost: ~,
browser_name: ~,
code_filepath: ~,
Expand Down Expand Up @@ -995,6 +997,7 @@ expression: "(&event.value().unwrap().spans, metrics.project_metrics)"
gen_ai_usage_output_tokens: ~,
gen_ai_usage_output_tokens_reasoning: ~,
gen_ai_response_model: ~,
gen_ai_request_model: ~,
gen_ai_usage_total_cost: ~,
browser_name: ~,
code_filepath: ~,
Expand Down Expand Up @@ -1293,6 +1296,7 @@ expression: "(&event.value().unwrap().spans, metrics.project_metrics)"
gen_ai_usage_output_tokens: ~,
gen_ai_usage_output_tokens_reasoning: ~,
gen_ai_response_model: ~,
gen_ai_request_model: ~,
gen_ai_usage_total_cost: ~,
browser_name: ~,
code_filepath: ~,
Expand Down Expand Up @@ -1482,6 +1486,7 @@ expression: "(&event.value().unwrap().spans, metrics.project_metrics)"
gen_ai_usage_output_tokens: ~,
gen_ai_usage_output_tokens_reasoning: ~,
gen_ai_response_model: ~,
gen_ai_request_model: ~,
gen_ai_usage_total_cost: ~,
browser_name: ~,
code_filepath: ~,
Expand Down
Loading